THE THING WE BUILT:

AI Receptionist

Last week, we explored how to respond when AI agents knock on our website’s front door. This week, we invite them in to our digital living room, encourage a candid business conversation, and notify us humans when we need to be brought into the loop.

But how’d we get here?

The rise of OpenClaw has us thinking about what happens when everyone co-works with their own agent. When agents can make calls, search the web, and even negotiate on your behalf. Agents that go out into the world with a task and come back with negotiated results, not just research.

What happens when those agents start showing up at your digital front door?

Not crawlers. Not bots scraping your site. Agents, empowered and sent by real humans, with necessary tasks, looking for actual help.

Right now, they can do some research, but often run into a wall, like a contact form or a chatbot.

But what if your front door had a receptionist?

Not a static FAQ. Not a chatbot with canned responses. An actual AI receptionist that can read a structured request, evaluate it against your real capabilities, and have an intelligent conversation about whether you're the right fit — before any human lifts a finger.

So we built a thing.

Our receptionist in action

How It Works

Here’s how it might play out in the real world:

1. A human tells their agent: “Our surgical robot’s touchscreen interface needs a redesign — surgeons are making errors under time pressure. We need someone who understands human factors.”

2. The agent breaks down the task — what it can research, investigate, decide, negotiate alone vs. which problems, questions, or tradeoffs need human expertise

3. It searches the web for providers

4. Most websites only have contact forms. One has an Agent Door — a machine-readable endpoint. You may remember the static version from last week. There’s more now.

5. The agent reads the manifest: what do the humans here actually provide? What should agents handle themselves?

6. It sends a structured negotiation request

7. The provider's AI receptionist evaluates: Is this a fit? Which capabilities matter here? Be honest.

8. The visiting agent translates everything back to plain language for its human

9. Both humans get notified — they decide whether to connect

Agents discover, qualify, and match. Humans decide and commit.

Notes from Building

This week, we built with OpenClaw. We heard the chatter and wanted to experiment ourselves. We set up OpenClaw, an autonomous AI agent, in a virtual Linux machine and used it both to help build the prototype and to act as our receptionist when the demo went live. When you ring the doorbell, OpenClaw responds based on the instructions we provided.

This happens in the background, so we made it visible. When you submit a task on our demo site, live API calls are made to Claude. Two independent AI agents, a visiting agent and Township’s AI receptionist, reason about your input in real time.

We built it as a split-screen experience so you can see behind the scenes. On the left, the conversation as a user would experience it. On the right, every request, every response, every model call as it happens.

As more AI moves into the background of our workflows, it becomes critical to understand what’s actually happening. We don’t just build the AI process. We build the visibility, control, and confidence around it.

This is a proof of concept. The protocol still needs authentication, logging, web search integration, and notification plumbing. But the core loop works today. Two agents. One protocol. Humans at both ends.

The question isn’t whether agents will start knocking on your door. They will. The question is whether anyone’s home when they do.

Built with ❤️ by Township

You’re receiving this because you signed up for Township’s “We Built a Thing” newsletter. We'll only send these when we, well, build things.